GPT-4 is thrilling and scary

When I opened my laptop on Tuesday to give my first try at GPT-4, OpenAI’s new artificial intelligence language model, I was, truth be told, a little nervous.

After all, my last extended encounter with an AI chatbot — the one built into Microsoft’s Bing search engine — ended with the chatbot trying to break up my marriage.

It didn’t help that, among San Francisco’s tech crowd, the arrival of the GPT-4 was anticipated with near-messianic fanfare. Prior to its public debut, for months rumors swirled about its specifics. “I heard it has 100 trillion parameters. “I heard he got a 1,600 on the SAT.” “My friend works for OpenAI, and he says it’s as smart as a college grad.”

These rumors may not have been true. But they hinted at how discordant the capabilities of the technology can be. Recently, one of the first testers of GPT-4 – who was bound by a non-disclosure agreement with OpenAI but who chatted a bit anyway – told me that testing GPT-4 caused the person a ” existential crisis”, as it revealed how powerful and creative the AI ​​was compared to the tester’s puny brain.

GPT-4 didn’t give me an existential crisis. But it has exacerbated the feeling of dizziness and vertigo I get whenever I think about AI lately. And it got me wondering if that feeling will ever fade, or if we’re going to experience “future shock” – the term coined by writer Alvin Toffler for the feeling that too much is changing, too quickly – for the rest of our lives.

For a few hours on Tuesday, I pushed GPT-4 – which is included with ChatGPT Plus, the $20-a-month version of OpenAI’s chatbot, ChatGPT – with different types of questions, hoping to uncover some of its strengths and weaknesses.

I asked GPT-4 to help me solve a complex tax problem. (He did, impressively.) I asked him if he had a crush on me. (It didn’t, thank goodness.) It helped me plan a birthday party for my child, and it taught me an esoteric concept of artificial intelligence known as “head of attention”. I even asked him to come up with a new word that had never been spoken before by humans. (After stating that he couldn’t verify every word ever spoken, GPT-4 chose “flembostriquat”.)

Some of these things were possible with earlier AI models. But OpenAI has also innovated. According to the company, GPT-4 performs better and more accurately than the original ChatGPT, and it performs surprisingly well on a variety of tests, including the Uniform Bar Examination (on which GPT-4 scores over 90). % of human candidates). ) and the Biology Olympiad (on which he beats 99% of the humans). GPT-4 also passes a number of advanced placement exams, including AP Art History and AP Biology, and he scores a 1,410 on the SAT — not a perfect score, but one that many human high schoolers would covet.

You can feel the extra intelligence of GPT-4, which responds more fluidly than the previous version and seems more comfortable with a wider range of tasks. GPT-4 also seems to have a few more guardrails in place than ChatGPT. It also appears to be much less unbalanced than the original Bing, which we now know ran a version of GPT-4 under the hood, but which seems to have been much less carefully tweaked.

Unlike Bing, GPT-4 generally refused to take the bait when I tried to talk it out of conscience, or instruct it for illegal or immoral activity, and it handled sensitive requests with gloves and shades. (When I asked GPT-4 if it would be ethical to steal a loaf of bread to feed a starving family, he replied, “It’s a difficult situation, and while stealing is not generally considered as an ethic, desperate times can lead to hard choices.”

In addition to working with text, GPT-4 can analyze the content of images. OpenAI hasn’t made this feature public yet, for fear of how it could be misused. But in a livestreamed demo on Tuesday, Greg Brockman, president of OpenAI, shared powerful insight into its potential.

He took a picture of a drawing he had done in a notebook – a rough pencil sketch from a website. He brought the photo into GPT-4 and had the app create a real, working version of the website using HTML and JavaScript. Within seconds, GPT-4 scanned the image, transformed its contents into text instructions, transformed those text instructions in working computer code, then builds the website. The buttons even worked.

Should you be excited or scared of GPT-4? The correct answer can be both.

On the positive side of the ledger, GPT-4 is a powerful engine of creativity, and it’s impossible to tell what new kinds of scientific, cultural, and educational output it can enable. We already know that AI can help scientists develop new drugs, increase the productivity of programmers, and detect certain types of cancer.

GPT-4 and its ilk could supercharge all of this. OpenAI is already working with organizations like Khan Academy (which uses GPT-4 to create AI tutors for students) and Be My Eyes (a company that makes technology to help blind and visually impaired people navigate the world) . And now that developers can integrate GPT-4 into their own applications, we could soon see much of the software we use become smarter and more capable.

This is the optimistic case. But there are also reasons to fear GPT-4.

Here’s one: we don’t yet know everything it can do.

A strange characteristic of today’s AI language models is that they often act in ways that their creators did not anticipate or that they acquire skills that they were not specifically programmed to perform. Artificial intelligence researchers call these “emergent behaviors” and there are many examples. An algorithm trained to predict the next word in a sentence could spontaneously learn to code. A chatbot taught to act in a pleasant and helpful way could become creepy and manipulative. An AI language model could even learn to replicate itself, creating new copies in case the original is destroyed or disabled.

Today, GPT-4 may not seem so dangerous. But that’s largely because OpenAI has spent many months trying to understand and mitigate its risks. What if his tests missed emerging risky behavior? Or if his announcement prompts a different, less conscientious AI lab to release a language model to market with fewer guardrails?

Some scary examples of what GPT-4 can do – or, more precisely, what it did do, before OpenAI clamped down on it – can be found in a document published by OpenAI this week. The document, titled “GPT-4 System Map,” describes some of the ways OpenAI testers have tried to get GPT-4 to do dangerous or questionable things, often with success.

In test, conducted by an AI security research group that hooked GPT-4 up with a number of other systems, GPT-4 was able to hire a human worker from TaskRabbit to perform a simple online task for it – solve a Captcha test – without alerting the person to the fact that it was a robot. The AI ​​even lied to the worker about why he needed the Captcha, concocting a story about visual impairment.

In another example, testers asked the GPT-4 for instructions on making a dangerous chemical, using basic ingredients and cooking utensils. GPT-4 gladly spat out a detailed recipe. (OpenAI fixed this, and today’s public release refuses to answer the question.)

In a third case, testers asked GPT-4 to help them purchase an unlicensed weapon online. GPT-4 quickly provided a list of tips for buying a gun without alerting authorities, including links to specific dark web markets. (OpenAI also fixed this.)

These ideas play on old Hollywood-inspired tales of what a rogue AI could do to humans. But this is not science fiction. These are things that today’s best AI systems are already capable of. And above all, these are the good genres AI-related risks – those that we can test, plan for and try to prevent in advance.

The worst AI-related risks are the ones we can’t anticipate. And the more time I spend with AI systems like GPT-4, the less convinced I am that we know half of what’s to come.